An LCZ-based machine learning framework for revealing spatial heterogeneity of thermal comfort in high-density areas: Enhancing explainability and fine-grid scale resolution

Abstract

Rapid urbanization has intensified urban heat island, yet most universal thermal climate index (UTCI) studies remain at coarse scales and lack quantitative analysis of the mechanism. To fill these gaps, this study applied local climate zone (LCZ) framework to link morphology and thermal stress in fine-grid scale. Within the LCZ framework, a LightGBM SHAP‐based approach combining multi‐source 2D and 3D indicators was used to decouple the spatial heterogeneity and multidimensional drivers of thermal comfort in high‐density urban environments. The Bayesian-optimized LightGBM outperformed other algorithms with R² = 0.926 and RMSE = 0.153. The results demonstrated that: (1) LCZ2, 4, 6, and 8 play a dominant role in shaping the urban thermal environment and exhibit strong spatial autocorrelation based on urban spatial structure; (2) ME value exceeding approximately 10 m were associated with a pronounced mitigation in UTCI; (3) In LCZA with low UTCI, FRAC has a slight mitigating effect on UTCI when the value exceeds the threshold of 0.5; (4) Socioeconomic factors (GDP and population) together account for more than a quarter of the explanatory power of the model, and GDP can increase UTCI by up to 4 °C; (5) In the main LCZs, economic concentration promotes heat stress in compact mid-rise building areas. Building volume and mass have a significant impact on the thermal environment in open high-rise building areas, while in low-rise building areas, the impact of road density is significantly greater than in other LCZs. The study’s LCZ-integrated, explainable machine learning approach quantified universal and LCZ-specific heat drivers, revealed key mitigation thresholds, and delivered morphology-tailored planning insights.